Does training on OBLIQ tasks bake in specific similarity notions?

antoine_chaffin · x · 2026-08-23

The author questions whether current model setups for OBLIQ tasks are simply training models for those specific benchmarks by baking in a particular notion of similarity, rather than learning a generalizable prompting mechanism. They express concern that this might be achieved easily via synthetic data but may fail on out-of-distribution tasks, comparing late interaction limits with approaches like Promptriever.

Original post →

More from Models

Models channel →